LongMemEval对话AI数据集|记忆评估数据集

LongMemEval是由腾讯AI实验室西雅图分部构建的长期记忆评估基准，包含500个高质量问题，覆盖信息提取、跨会话推理、时间推理、知识更新和拒绝回答五类核心能力。数据基于多轮任务对话生成，支持灵活配置历史长度，提供约115k和1.5M tokens的标准设置。采用属性控制流水线确保对话连贯性与可扩展性，适用于评估聊天助手在长程交互中的记忆性能。LongMemEval is a long-term memory evaluation benchmark developed by the Seattle Branch of Tencent AI Lab. It contains 500 high-quality questions covering five core capabilities: information extraction, cross-session reasoning, temporal reasoning, knowledge update, and refusal to answer. The dataset is generated from multi-turn task dialogues, supports flexible configuration of conversation history length, and provides two standard settings with approximately 115k and 1.5M tokens respectively. An attribute-controlled pipeline is adopted to ensure dialogue coherence and scalability, making it suitable for evaluating the memory performance of chat assistants in long-range conversational interactions.

搜集汇总

背景与挑战

背景概述

LongMemEval是由腾讯AI实验室构建的长期记忆评估基准，包含500个高质量问题，覆盖信息提取、跨会话推理等五类核心能力，用于测试聊天助手在长程交互中的记忆性能。数据基于多轮任务对话生成，支持灵活配置历史长度，并提供标准token设置，采用属性控制流水线确保对话连贯性与可扩展性。